Deep metric learning-based side-channel analysis with improved robustness and efficiency

Deep metric learning-based side-channel analysis with improved robustness and efficiency

연구 분야: Verification

학회: Applied Intelligence

Side-channel analysis (SCA) is one of the widely studied approaches for assessing vulnerabilities in cryptographic algorithm implementations. Existing deep learning (DL)-based SCA approaches are commonly dataset-specific, and their attack performance heavily depends on optimal hyperparameters and effective neural network architectures. Searching such hyperparameters and architectures could be very time-consuming. In addition, traditional machine learning (ML)-based SCA methods often require manual feature engineering, leading to information loss and limiting attack performance. To address these challenges, we propose a profiled SCA model based on deep metric learning (DML) with template attacks (TA). This novel approach improves dataset generalization, enhances feature extraction, and reduces the reliance on hyperparameters. Specifically, a normalized lifted structured (NLS) loss is designed for the proposed attack model. Then, a label-informed hybrid distance is subtly integrated into the model to enhance the model’s ability for capturing relationships between embeddings and labels, thereby improving the attack performance and robustness. Next, a similarity learning method is designed by evaluating all pairwise distances within a mini-batch, reducing sensitivity to triplet selection and improving training efficiency. Experimental results show that the proposed model significantly outperforms the state-of-the-art DL-based SCA methods. It achieves attack performance improvements of up to 50.0% and an average improvement of 37.9% on public datasets, while being 30.8% faster in network training. Comprehensive evaluations show that the proposed model provides high efficiency, robust performance, and strong generalization across diverse datasets and leakage models.